Invisible Poison: A Blackbox Clean Label Backdoor Attack to Deep Neural Networks

This paper reports a new clean-label data poisoning backdoor attack, named Invisible Poison, which stealthily and aggressively plants a backdoor in neural networks. It converts a regular trigger to a noised trigger that can be easily concealed inside images for training NN, with the objective to plant a backdoor that can be later activated by the trigger. Compared with existing data poisoning backdoor attacks, this newfound attack has the following distinct properties. First, it is a blackbox attack, requiring zero-knowledge of the target model. Second, this attack utilizes "invisible poison" to achieve stealthiness where the trigger is disguised as `noise', and thus can easily evade human inspection. On the other hand, this noised trigger remains effective in the feature space to poison training data. Third, the attack is practical and aggressive. A backdoor can be effectively planted with a small amount of poisoned data and is robust to most data augmentation methods during training. The attack is fully tested on multiple benchmark datasets including MNIST, Cifar10, and ImageNet10, as well as application specific data sets such as Yahoo Adblocker and GTSRB. Two countermeasures, namely Supervised and Unsupervised Poison Sample Detection, are introduced to defend the attack.